Kernelled Naive Bayes Using a Balanced Dataset for Accurate Classification of the Material Toxicity

In this work, a new multiclass classification approach was employed in the QSAR model to assess chemical toxicity prediction through handling the imbalanced dataset as the critical preprocessing step in the training dataset. Various classifiers of the decision tree, KNN, naive Bayes, kernelled naive Bayes, and SVM and two distinct acute aquatic toxicity datasets towards Daphnia Magna and Fathead Minnow Fish were used to evaluate the generality of the approach. The quantitative response (LC50) was discretized into ten bins. Imbalanced dataset classification leads to a high level of errors since the classifier tends to learn from the majority class more than the minority class. Each training dataset was specified by different weights related to the class population. These datasets were then bootstrapped based on their weights to convert the imbalanced dataset into a balanced one. This approach enhanced the accuracy of classification of material toxicity dramatically (up to 99%). Balanced dataset classification had high overall accuracy when correlated attributes were removed. Therefore, fewer attributes are sufficient to predict material toxicity.The overall accuracy improvement of the decision tree, K-NN, naive Bayes, kernelled naive Bayes, and SVM for the Daphnia Magna dataset after balancing the data set are 58.03%, 55.08%, 9.09%, 72.48%, and 53.05%, respectively.